A Control-Theoretic Approach to Inertial SLAM

Abstract

A team of Un-manned Aerial Vehicles (UAVs) is tasked to explore an unknown environment and to map the features they find, but must do so without the use of infrastructure based localisation systems such as the Global Positioning System (GPS), or any a-prior terrain data. The UAVs navigate using a statistical estimation technique known as Simultaneous Localisation And Mapping (SLAM) which allows for the simultaneous estimation of the location of the UAV as well as the location of the features it sees. SLAM offers a unique approach to vehicle localisation with potential applications including planetary exploration, or when GPS is denied (for example under intentional GPS jamming, or applications where GPS signals cannot be reached), but more importantly can be used to augment already existing systems to improve robustness to navigation failure. One key requirement for SLAM to work is that it must re-observe features, and this has two effects: firstly, the improvement of the location estimate of the feature; and secondly, the improvement of the location estimate of the platform because of the statistical correlations that link the platform to the feature. So our UAV has two options; should it explore more unknown terrain to find new features, or should it revisit known features to improve localisation quality. Additionally, it is known that the maneuvers the agent takes during feature observations affects the accuracy in localisation estimates and hence the accuracy of the constructed map. This thesis is concerned with studying the interaction and tight coupling between the processes of SLAM and motion planning/control of an autonomous intelligent agent. We focus on inertial-sensor based SLAM due to its applicability to several different vehicle modalities. Architectures for inertial SLAM are presented for both global and local-scale environments, with the estimation of inertial sensor biases, with both range/bearing and bearing-only terrain sensors and for both single and multiple vehicles. The aim is to demonstrate a valid theoretic implementation which is used as the foundation for a study of the algorithm. We begin by studying the observability properties of the inertial SLAM algorithm, focussing on the connection between vehicle dynamic maneuvers and the observability of the equations. We then consider the problem of ‘active SLAM’ , where the agent makes intelligent control decisions in order to exploit the coupling between SLAM accuracy and agent motion. The analysis is then extended to the multi-agent case and several control strategies are demonstrated. Simulation results of implementations of the SLAM algorithm, maneuver analysis and both single and multi-vehicle active SLAM architectures are presented using a six-degree of freedom, multi-UAV simulator.